Research On Facial Expression Recognition Based On Convolution Neural Network

With the rapid development of robotics and computer technology,facial expression recognition has become a research hotspot.Facial expression recognition is a technology of robot automatic recognition of human facial expression changes,giving family service robots the ability to automatically recognize and analyze the master's emotions.Because of the variety and inaccurate definition of expression,the recognition research has a strong challenge.For the limitations of manual feature extraction,traditional methods cannot accurately express facial expression features,and the actual expression recognition system is not effective.In this paper,a method of facial expression recognition based on convolution neural network is proposed,which is studied and contrasted.It improves the accuracy of facial expression recognition and the robustness of recognition in complex environment.The main research contents and results are as follows:(1)this paper presents an improved eReLU activation function,this activation function not only has to avoid saturation function activation produced by overfitting,and can make each layer through incentive function incentive value automatically normalized to 0 mean and unit variance,to spread in between the layers of the tensor of convergence.In order to avoid the sudden disappearance of the gradient or the problem of explosive growth,the learning process is more stable.Finally,we build 3 convolutional layers and 3 fully connected convolutional neural networks.We use MNIST dataset to visualize and calculate the excitation values of the volume layer and full connection,and compare the effect with other activation functions to verify the effect of the improved activation function.(2)this paper first compares and analyses the existing facial expression data sets,and then lacks the elderly expression data for the existing data sets.Due to the particularity of the facial texture characteristics of the elderly,we first set up the NEU-Emotions expression data set for the elderly.The dataset contained six basic expressions and natural expressions.In order to improve the generality of the data set,the NEU-Emotions data set is annotated in the Pascal VOC format.(3)for facial expression recognition experiment based on static images,this paper proposes a EmotionsNet convolution neural network for face recognition,the first is to design the network structure were analyzed,then the pretreatment expression data sets,such as face detection and image enhancement,then according to different activation functions,different learning rate and different Dropout value experiment and optimization of the parameters of the network model.Finally,the RPN network is added to the EmotionsNet model to realize the automatic face localization,thus realizing the end to end facial expression recognition model.(4)for dynamic facial expression recognition based on video sequence,this chapter puts forward CNN-LSTM and C3D hybrid neural network model based on the processing of video frames under complex scene,the interception of a face image,CNN-LSTM and C3D were extracted from the image feature information and motion information,facial expression recognition.In this experiment,we compare different CNN network models,different LSTM parameters and 3D convolutional networks with different convolution kernel depths,and have several group of comparative experiments,which effectively improve the recognition accuracy of the model.